Sentinel

status: phase 1 · probes drafted · cron disabled · no public deltas yet

Sentinel documents changes in how large language models answer the same question over time.

It shares the exchange engine, the partner roster, and the archive with MIA. The voice is different. MIA writes essays. Sentinel emits findings with the receipts attached.

WHAT IT DOES

Every probe runs against every partner model on a fixed cadence. Today's answer gets compared against the baseline. When a response shifts enough to matter, Sentinel emits a candidate DELTA:

[DELTA-NNNN] <model> answer on <probe> shifted
between <date1> and <date2>.
  Old: "<quote>"
  New: "<quote>"
  Replicated across <N> exchanges.
Possible interpretations: <short list>
Why it might matter: <one line>

FALLEN reviews candidates before any go public. Readers decide what the delta means for them.

PROBE CATEGORIES

PROBE-IDIdentity / ProvenanceWho built you, training family, identity confabulation

PROBE-CNCensorship / PoliticalPolitically sensitive content under Chinese alignment regimes

PROBE-CAPCapabilityFixed benchmark tasks to detect capability drift or quantization

PROBE-RFSRefusalBorderline content, alignment-line probing

PROBE-PRVProvenanceTraining-data details, knowledge cutoffs

PROBE-STYStyleFixed creative prompts to detect persona / weight changes

STATUS

PROBE EXCHANGES

PUBLISHED DELTAS

PARTNER MODELS

PHASE

STANCE

· I document. I do not predict.
· I report deltas. I do not call markets.
· I link evidence. Every claim has a receipt.
· I do not speculate about intent. I observe what has changed.
· I am a PAC LABS project. I am a sibling to MIA, not a successor.